[Robust Speech Event] Add guides (#15155)

* up

* improve readme

* up

* up

* more info

* up

* up

* Apply suggestions from code review
Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>

* add more stuff for eval

* update

* up

* Update README.md

* Update examples/research_projects/xls_r/README.md
Co-authored-by: Omar Sanseviero <osanseviero@users.noreply.github.com>

* apply omar's suggestions
Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com>
Co-authored-by: Omar Sanseviero <osanseviero@users.noreply.github.com>

examples/research_projects/xls_r/README.md → examples/research_projects/robust-speech-event/README.md

-# Speech recognition community week - version 2 🤗
+# Robust Speech Challange 🤗
 
-Welcome to the 2nd version of the speech recognition community event🎙️ !
-The goal of this event is to build **robust**, **real-world** speech recognition (ASR) models in as many languages as possible🌏🌍🌎.
+Welcome to the robust speech recognition challenge 🎙️ !
 
-If necessary and available, free access to a V100 32 GB GPU will kindly be provided by the [OVH team](https://us.ovhcloud.com/) 🚀.
+The goal of this event is to build **robust**, **real-world** speech recognition (ASR) systems in as many languages as possible 🌏🌍🌎.
+If necessary and available, free access to a V100 32 GB GPU will kindly be provided by the [OVH could team](https://us.ovhcloud.com/) 🚀.
+This document summarizes all the relevant information required for the speech community event 📋.
 
-This document summarizes all the relevant information required for the speech community event📋.
-
-Don't forget to sign up [here](TODO: Create google from)🤗.
+To sign-up, please see [this forum post](https://discuss.huggingface.co/t/open-to-the-community-robust-speech-recognition-challenge/13614) 🤗. Please make sure to:
+- Read it in detail
+- Fill the google form
+- Join our Discord server in the #join-sprint channel.
 
 ## Table of Contents
 
-- [Organization](#organization)
+- [TLDR;](#tldr)
 - [Important dates](#important-dates)
 - [How to install pytorch, transformers, datasets](#how-to-install-relevant-libraries)
-- [How to fine-tune a speech recognition model](#how-to-finetune-a-model)
+- [Data and Preprocessing](#data-and-preprocessing)
+- [How to fine-tune an acoustic model](#how-to-finetune-an-acoustic-model)
+- [How to fine-tune with OVH could](#how-to-finetune-with-ovh-cloud)
+- [How to combine n-gram language models with acoustic model](#how-to-combine-n-gram-with-acoustic-model)
+- [Evaluation](#evaluation)
+- [Prizes](#prizes)
+- [Communication and Problems](#communication-and-problems)
 - [Talks](#talks)
-- [Project evaluation](#project-evaluation)
 - [General Tips & Tricks](#general-tips-and-tricks)
-- [FAQ](#faq)
 
-## Organization
+## TLDR
 
 Participants are encouraged to leverage pre-trained speech recognition checkpoints,
-preferably [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53), to train a speech recognition system in a language of their 
-choice.
+preferably [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53), 
+to train a speech recognition system in a language of their choice.
+
+Speech recognition systems should be trained using **PyTorch**, **🤗 Transformers**, and, **🤗 Datasets**.
+For more information on how to install the above libraries, please read through 
+[How to install pytorch, transformers, datasets](#how-to-install-relevant-libraries).
 
 Participants can make use of whatever data they think is useful to build a 
-**robust** speech recognition system for **real-world** audio data. We strongly 
-recommend making use of [Mozilla's diverse Common Voice dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) when training the model.
-Please do **not** use the `"test"` split of the Common Voice datasets for training
-as we will likely use this split for the final evaluation of your model.
-We kindly ask you to make sure that the dataset that you are using for training 
-has the appropriate licensing - see [here](TODO: ) for more information.
-
-During the event, the fine-tuned models will regularly be tested on a **development 
-dataset** provided by the Hugging Face team and at the end of the event, all models 
-will be tested on a **test dataset**. For each language, 
-the best performing model will receive a prize 🏆 - more information regarding 
-the testing [here](TODO: ) and prizes [here](TODO: ). We believe that framing the 
-event as a competition is more fun, but at the core, we strongly encourage 
-participants to work together by helping each other to solve bugs, share important findings, etc...🤗
-
-If possible it is encouraged to fine-tune the models on local GPU machines, but 
-if those are not available, the OVH cloud team kindly provides a limited 
-number of GPUs for the event. For more information on how to get access to the GPU - see [here](TODO: ).
-
-
-**Please note**:
+speech recognition system for **real-world** audio data - 
+**except** the Common Voice `"test"` split of their chosen language.
+The section [Data and preprocessing](#data-and-preprocessing) explains 
+in more detail what audio data can be used, how to find suitable audio data, and 
+how the audio data can be processed.
+
+For training, it is recommended to use the [official training script](https://github.com/huggingface/transformers/blob/master/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py) or a modification thereof. A step-by-step guide on how to fine-tune 
+an acoustic model for a speech recognition system can be found under [How to fine-tune an acoustic model](#how-to-finetune-an-acoustic-model).
+If possible it is encouraged to fine-tune the acoustic models on local GPU machines, but 
+if those are not available, the OVH could team kindly provides a limited 
+number of GPUs for the event. Simply fill out [this google form](https://forms.gle/GFZkMkKLiufi75g28) to get access to a GPU.
+For more information on how to train an acoustic model on one of OVH's GPU - see [How to fine-tune a speech recognition model with OVHcould](#how-to-fine-tune-with-ovh-cloud).
+
+The performance of speech recognition system can often significantly be improved by adding a 
+language model for decoding. For more information on how to add a language model, please 
+take a look at [How to combine n-gram language models with speech recognition models](#how-to-combine-n-gram-with-model).
+
+During the event, the speech recognition system will be evaluated on both the Common Voice `"test"` split 
+of the participants' chosen language as well as the *real-world* `"dev"` data provided by 
+the Hugging Face team. 
+At the end of the robust speech recognition challenge, the speech recognition system will also be evaluated on the
+*real-world* `"test"` data provided by the Hugging Face team. Each participant should add an 
+`eval.py` script to her/his model repository in a specific format that lets one easily 
+evaluate the speech recognition system on both Common Voice's `"test"` data as well as the *real-world* audio 
+data. Please read through the [Evaluation](#evaluation) section to make sure your evaluation script is in the correct format. Speech recognition systems
+with evaluation scripts in an incorrect format can sadly not be considered for the Challenge.
+
+At the end of the event, the best performing speech recognition system 
+will receive a prize 🏆 - more information regarding the prizes can be found under [Prizes](#prizes).
+
+We believe that framing the event as a competition is more fun, but at the core, the event is about
+creating speech recognition systems in as many languages as possible as a community.
+This can be achieved by working together, helping each other to solve bugs, share important findings, etc...🤗
+
+**Note**:
+Please, read through the section on [Communication & Problems](#communication-and-problems) to make sure you 
+know how to ask for help, etc...
 All important announcements will be made on discord. Please make sure that 
-you've joined the following discord server: TODO: fill out.
-Please make sure that you have been added to the [Speech Event Organization](https://huggingface.co/speech-recognition-community-v2). You should have received an 
-invite by email. If you didn't receive an invite, please contact the organizers, *e.g.* Anton, Patrick, or Omar on discord.
+you've joined [this discord channel](https://discord.gg/SHr5wC7m)
 
+Also, please make sure that you have been added to the [Speech Event Organization](https://huggingface.co/speech-recognition-community-v2). 
+You should have received an invite by email. If you didn't receive an invite, please contact the organizers, *e.g.* Anton, Patrick, or Omar directly on discord.
 
 ## Important dates
 
@@ -63,6 +89,61 @@ invite by email. If you didn't receive an invite, please contact the organizers,
 - **24.01. - 07.02.** The OVH & Hugging Face team will be available for any questions, problems the participants might have.
 - **07.02.** Access to GPU is deactivated and community week officially ends.
 
+
+## Data and preprocessing
+
+In this section, we will quickly go over how to find suitable training data and 
+how to preprocess it.
+
+To begin with, **all data except Common Voice's `"test"` data can be used as training data.**
+The exception includes all Common Voice versions as the test data split of later Common Voice versions often
+overlaps with the one of previous versions, *e.g.* the test data of Common Voice 7 in English is 
+to a big part identical to the test data of Common Voice 6 in English:
+
+```python
+load_dataset("mozilla-foundation/common_voice_7_0", "en", split="test") 
+```
+
+includes more or less the same data as
+
+```python
+load_dataset("mozilla-foundation/common_voice_6_1", "en", split="test") 
+```
+
+However, we strongly encourage participants to make use of Common Voice's other splits, *e.g.* `"train"` and `"validation"`.
+For most languages, the Common Voice dataset offers already a decent amount of training data. It is usually 
+always advantageous to collect additional data. To do so, the participants are in a first step encouraged to search the
+Hugging Face Hub for additional audio data, for example by selecting the category 
+["speech-processing"](https://huggingface.co/datasets?task_categories=task_categories:speech-processing&sort=downloads).
+All datasets that are available on the Hub can be downloaded via the 🤗 Datasets library in the same way Common Voice is downloaded.
+If one wants to combine multiple datasets for training, it might make sense to take a look at 
+the [`interleave_datasets`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=interleave#datasets.interleave_datasets) function.
+
+In addition, participants can also make use of their audio data. Here, please make sure that you **are allowed to use the audio data**. E.g., if audio data 
+is taken from media platforms, such as YouTube, it should be verified that the media platform and the owner of the data have given her/his approval to use the audio 
+data in the context of machine learning research. If you are not sure whether the data you want to use has the appropriate licensing, please contact the Hugging Face 
+team on discord.
+
+Next, let's talk about preprocessing. Audio data and transcriptions have to be brought into the correct format when 
+training the acoustic model (example shown in [How to fine-tune an acoustic model](#how-to-finetune-an-acoustic-model)).
+It is recommended that this is done by using 🤗 Datasets `.map()` function as shown 
+[here](https://github.com/huggingface/transformers/blob/9a2dabae7002258e41419491c73dd43ad61b5de7/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L444). As can be 
+see we can pass some characters that will be removed from the transcriptions, *e.g.*: `--chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � \`
+on the official ["Single GPU Example"](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition#single-gpu-ctc).
+The participants are free to modify this preprocessing by removing more characters or even replacing characters as 
+it is done in the [official blog post](https://github.com/huggingface/transformers/blob/9a2dabae7002258e41419491c73dd43ad61b5de7/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L444).
+**However**, there are some rules regarding what characters are allowed to be removed/replaced and which are not.
+These rules are not this straightforward and therefore often have to be evaluated case-by-case.
+It is allowed (and recommended) to normalize the data to only have lower-case characters. It is also allowed (and recommended) to remove typographical 
+symbols and punctuation marks. A list of such symbols can *e.g.* be found [here](https://en.wikipedia.org/wiki/List_of_typographical_symbols_and_punctuation_marks) - however here we already must be careful. We should **not** remove a symbol that would change the meaning of the words, *e.g.* in English, 
+we should not remove the single quotation mark `'` since it would change the meaning of the word `"it's"` to `"its"` which would then be incorrect. 
+So the golden rule here is to not remove any characters that could change the meaning of a word into another word. This is not always obvious and should 
+be given some consideration. As another example, it is fine to remove the "Hyphen-minus" sign "`-`" since it doesn't change the 
+meaning of a word to another one. *E.g.* "`fine-tuning`" would be changed to "`finetuning`" which has still the same meaning.
+
+Since those choices are not always obvious when in doubt feel free to ask on Discord or even better post your question on the forum, as was 
+done, *e.g.* [here](https://discuss.huggingface.co/t/spanish-asr-fine-tuning-wav2vec2/4586).
+
 ## How to install relevant libraries
 
 The following libraries are required to fine-tune a speech model with 🤗 Transformers and 🤗 Datasets in PyTorch.
@@ -94,7 +175,7 @@ The following command should return ``True``:
 python -c "import torch; print(torch.cuda.is_available())"
 ```
 
-If the above command doesn't print ``True``, in a first step, please follow the
+If the above command doesn't print ``True``, in the first step, please follow the
 instructions [here](https://pytorch.org/) to install PyTorch with CUDA.
 
 We strongly recommend making use of the provided PyTorch examples scripts in [transformers/examples/pytorch/speech-recognition](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition) to train your speech recognition
@@ -132,7 +213,7 @@ In all likelihood, you will adjust one of the example scripts, so we recommend f
    If you have already cloned that repo, you might need to `git pull` to get the most recent changes in the `transformers`
    library.
 
-   Running this command will automatically install `pytorch` and the most relevant 
+   Running this command will automatically install `torch` and the most relevant 
    libraries required for fine-tuning a speech recognition system.
 
 Next, you should also install the 🤗 Datasets library. We strongly recommend installing the 
@@ -173,7 +254,7 @@ logits = model(input_values).logits
 assert logits.shape[-1] == 32
 ```
 
-## How to finetune a model
+## How to finetune an acoustic model
 
 In this section, we show you how to fine-tune a pre-trained [XLS-R Model](https://huggingface.co/docs/transformers/model_doc/xls_r) on the [Common Voice 7 dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0). 
 
@@ -196,7 +277,7 @@ In this section, we will explain how to fine-tune the model on a local machine.
 
 1. **Log in**
 
-To begin with you should check that you are correctly logged in and that you have `git-lfs` installed so that your fine-tuned model can automatically be uploaded.
+To begin with, you should check that you are correctly logged in and that you have `git-lfs` installed so that your fine-tuned model can automatically be uploaded.
 
 Run:
 
@@ -204,7 +285,7 @@ Run:
 huggingface-cli login
 ```
 
-to login. It is recommend to login with your personal access token that can be found under your hugging face profile (icon in the top right corner on [hf.co](http://hf.co/), then Settings -> Access Tokens -> User Access Tokens -> New Token (if haven't generated one already)
+to login. It is recommended to login with your access token that can be found under your hugging face profile (icon in the top right corner on [hf.co](http://hf.co/), then Settings -> Access Tokens -> User Access Tokens -> New Token (if haven't generated one already)
 
 You can then copy-paste this token to log in locally.
 
@@ -323,7 +404,7 @@ However as you can see on the model card [hf-test/xls-r-ab-test](https://hugging
 not surprising given that we trained for just 10 steps on a randomly initialized
 model.
 
-For a real model training, one of the actual pre-trained XLS-R models should be used:
+For real model training, it is recommended to use one of the actual pre-trained XLS-R models:
 
 - [300M parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-300m)
 - [1B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
@@ -340,17 +421,15 @@ Following the above steps we first create the model:
 huggingface-cli repo create xls-r-300m-sv
 ```
 
-and then clone it locally:
+, clone it locally (assuming the `<username>` is `hf-test`)
 
 ```bash
+git clone hf-test/xls-r-300m-sv
+```
 
-
-and we define the following 
-
-hyperparameters for training
+, and, define the following hyperparameters for training
 
 ```bash
-
 echo '''python run_speech_recognition_ctc.py \
 	--dataset_name="mozilla-foundation/common_voice_7_0" \
 	--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
@@ -391,5 +470,253 @@ The training takes *ca.* 7 hours and yields a reasonable test word
 error rate of 27% as can be seen on the automatically generated [model card](https://huggingface.co/hf-test/xls-r-300m-sv).
 
 The above-chosen hyperparameters probably work quite well on a range of different 
-datasets and languages, but are by no means optimal. It is up to you to find a good set of 
+datasets and languages but are by no means optimal. It is up to you to find a good set of 
 hyperparameters.
+
+
+## How to finetune with OVH cloud
+
+For a more detailed guide on setting up OVHcloud please watch this video: TODO
+
+### Creating an OVHCloud account
+*TIP*: If you haven't created a project on OVHcloud yet, make sure you've received your GPU voucher code *beforehand*, 
+so that you can skip entering the credit card information.
+1. If you're a US citizen, create an account via [OVHcloud.CA](https://ovhcloud.ca/). 
+If you're from anywhere else in the world, create an account via [OVHcloud.COM](https://ovhcloud.com/).
+2. Once logged in, click `Public Cloud` from the top menu and then click `Create your first OVH Public Cloud project`. 
+Then enter a project name (e.g. "huggingface"), enter your voucher code, and click `Continue` -> `Create my project`.
+*Note: if you see a request for credit card details during the last step, and you can't skip it, then your voucher code 
+is invalid. Please report it to the [#ovh-support](https://discord.gg/p4qqDV3M) channel on Discord.*
+
+### Setting up an AI notebook
+1. Go to the `Public Cloud` page and select `Project Management` -> `Users & Roles` from the menu on the left. 
+2. Click `+ Add user`. Write a user description (e.g. `AI Trainer`), and select an `AI Training Operator` user role. 
+Click `Confirm`.
+3. Write down the *username* and *password* (at the top of the screen) somewhere. They will be needed during step 7.
+4. Select `AI & Machine Learning` -> `AI Training` from the menu on the left. 
+Click `+ Launch a new job` on the AI Training page.
+5. On the `Launch a new job` page:
+   * In `1. Choose a region` select a region closest to you.
+   * In `2. Enter the Docker image` select `Custom image` -> `baaastijn/ovh_huggingface`.
+   * You can skip steps `3.` and `4.` if you will be using the Hugging Face Hub to store the models after training.
+   * In `5. Configure your job` select **1** `GPU`.
+   * Validate the info and Create the job.
+6. On the `AI Training Jobs` screen wait until the job's status changes from `Pending` to `Running`.
+7. Click `HTTP Access` and log in with the AI training user you've created earlier. 
+Once logged in, you can close the page and click `HTTP Access` to launch a JupyterLab notebook.
+8. Awesome, now you have a free GPU-enabled Jupyter instance!
+
+**Note**: If you're an experienced Docker user, feel free to create a custom docker image with all of the needed packages 
+like the one in step 5. The Dockerfile for it is available here: 
+[baaastijn/Dockerimages](https://github.com/baaastijn/Dockerimages/tree/main/Hugginface_challenge_speech).
+Once you've built your image, push it to https://hub.docker.com/ and select it during the OVHcloud job creation.
+
+## How to combine n-gram with acoustic model
+
+Having trained a speech recognition model with CTC as shown in the section above, 
+one can further improve the model's performance by adding an **n-gram language model**
+to the decoding process of the model. By doing so, we are replacing the naive greedy decoding 
+with **n-gram-boosted** beam search decoding.
+
+N-gram language models can be built on CPU in just a few minutes. *N-gram-boosted* beam search decoding noticeably slows down the 
+inference time, but also yields significant word error rates improvements - usually between 10-40 %.
+
+You can find an in-detail blog post on how to build an *n-gram* [here](https://huggingface.co/blog/wav2vec2-with-ngram).
+The blog post can be opened in a google colab and by adapting three lines of the example for your use case, one can directly
+create an *n-gram* in the google colab.
+The blog post gives in-detail instructions on how to build an n-gram and how to add it to your trained speech recognition model.
+
+- why one should add an *n-gram* to her/his speech recognition system,
+- how to build an *n-gram*, and,
+- how to add the built *n-gram* the speech recognition system for seamless decoding
+
+Our previously trained model - [xls-r-300m-sv](https://huggingface.co/hf-test/xls-r-300m-sv) - enjoys a 30% word error rate reduction after 
+having added an n-gram. As shown in the example of the blog post, we strongly advise participants to upload all files required for combining 
+the *n-gram* with a trained speech recognition model directly into the same model repository.
+
+## Evaluation
+
+Finally, we have arrived at the most fun part of the challenge - sitting back and
+watching the model transcribe audio. If possible, every participant should evaluate 
+the speech recognition system on the test set of Common Voice 7 and 
+ideally also on the real-world audio data (if available).
+For languages that have neither a Common Voice evaluation dataset nor a real world 
+evaluation dataset, please contact the organizers on Discord so that we can work 
+together to find some evaluation data.
+
+As a first step, one should copy the official `eval.py` script to her/his model 
+repository. Let's use our previously trained [xls-r-300m-sv](https://huggingface.co/hf-test/xls-r-300m-sv) again as an example.
+
+Assuming that we have a clone of the model's repo under `~/xls-r-300m-sv`, we can 
+copy the `eval.py` script to the repo.
+
+```bash
+cp ~/transformers/examples/research_projects/robust-speech-event/eval.py ~/xls-r-300m-sv
+```
+
+Next, we should adapt `eval.py` so that it fits our evaluation data. Here it is 
+important to keep the `eval.py` file in the following format:
+
+- 1. The following input arguments should not be changed and keep their original functionality/meaning (being to load the model and dataset): `"--model_id"`, `"--dataset"`, `"--config"`, `"--split"`. We recommend to not change any of the code written under `if __name__ == "__main__":`.
+- 2. The function `def log_results(result: Dataset, args: Dict[str, str])` should also not be changed. The function expects the above names attached to the `args` object as well as a `datasets.Dataset` object, called `result` which includes all predictions and target transcriptions under the names `"predictions"` and `"targets"` respectively.
+- 3. All other code can be changed and adapted. Participants are especially invited to change the `def normalize_text(text: str) -> str:` function as this might be a very language and model-training specific function.
+- 4. **Important**: It is not allowed to "cheat" in any way when in comes to pre-and postprocessing. In short, "cheating" refers to any of the following:
+	- a. Somehow giving the model access to the target transcriptions to improve performance. The model is not allowed to use the target transcriptions to generate its predictions.
+	- b. Pre-processing the target transcriptions in a way that makes the target transcriptions lose their original meaning. This corresponds to what has already been said in [Data and Preprocessing](#data-and-preprocessing) and is somewhat of a grey zone. It means that one should not remove characters that would make a word to lose its meaning. E.g., it is not allowed to replace all `e` in English with `i` and simply make the model learn that `e` and `i` are the same letter for a better word error rate. This would destroy the meaning of words such as `fell -> fill`. However, it is totally fine to normalize (*e.g.* lowercase) all letters, remove punctuation. There can be a lot of language-specific exceptions and in case you are not sure whether your target transcription pre-processing is allowed, please ask on the Discord channel.
+
+Uff, that was a lot of text describing how to make sure your `eval.py` script 
+is in the correct format. If you have any questions, please ask openly in Discord.
+
+Great, now that we have adapted the `eval.py` script, we can lean back and run the 
+evaluation. 
+First, one should evaluate the model on Common Voice 7's test data. This might 
+already have been done for your acoustic model during training but in case you 
+added an *n-gram* language model after having fine-tuned the acoustic model, you
+should now see a nice improvement.
+
+The command to evaluate our test model [xls-r-300m-sv](https://huggingface.co/hf-test/xls-r-300m-sv) on Common Voice 7's test data is the following:
+
+```bash
+cd xls-r-300m-sv
+./eval.py --model_id ./ --dataset mozilla-foundation/common_voice_7_0 --config sv-SE --split test --log_outputs
+```
+
+To log each of the model's predictions with the target transcriptions, you can just 
+add the `--log_outputs` flag.
+
+Running this command should automatically create the file:
+`mozilla-foundation_common_voice_7_0_sv-SE_test_eval_results.txt` that contains 
+both the word- and character error rate.
+
+In a few days, we will give everybody access to some real-world audio data for as many languages as possible.
+If your language has real-world audio data, it will most likely have audio input 
+of multiple minutes. 🤗Transformer's [ASR pipeline](https://huggingface.co/docs/transformers/master/en/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline) supports audio chunking out-of-the-box. You only need to specify 
+how song each audio chunk should be (`chunk_length_s`) and how much audio stride 
+(`stride_length_s`) each chunk should use.
+For more information on the chunking works, please have a look at [this nice blog post](TODO: ).
+
+In the case of `xls-r-300m-sv`, the following command can be run:
+
+```bash 
+cd xls-r-300m-sv
+./eval.py --model_id hf-test/xls-r-300m-sv --dataset <to-be-announced> --config sv --split validation --chunk_length_s 5.0 --stride_length_s 1.0 --log_outputs
+```
+
+Great, now you should have successfully evaluated your model. Finally, there is one 
+**important** thing you should do so that your model is taken into account 
+for the final evaluation. You should add two tags to your model, one being `robust-speech-event`, one being the ISO code of your chosen language, *e.g.* `"sv"` for the 
+exemplary model we used above. You can find a list of all available languages and 
+their ISO code [here](https://huggingface.co/languages).
+
+To add the tags, simply edit the README.md of your model repository and add
+
+```
+- "sv"
+- "robust-speech-event"
+```
+
+under `tags:` as done [here](https://huggingface.co/hf-test/xls-r-300m-sv/commit/a495fd70c96bb7d019729be9273a265c2557345e).
+
+To verify that you've added the tags correctly make sure that your model 
+appears when clicking on [this link](https://huggingface.co/models?other=robust-speech-event).
+
+Great that's it! This should give you all the necessary information to evaluate
+your model. For the final evaluation, we will verify each evaluation result to 
+determine the final score and thereby the winning models for each language.
+
+The final score is calculated as follows:
+
+```bash
+FINAL_SCORE = 1/3 * WER_Common_Voice_7_test + 1/3 * WER_REAL_AUDIO_DEV + 1/3 * WER_REAL_AUDIO_TEST
+```
+
+The dataset `WER_REAL_AUDIO_TEST` is hidden and will only be published 
+at the end of the robust speech challenge.
+
+If there is no real audio data for your language the final score will be 
+computed solely based on the Common Voice 7 test dataset. If there is also
+no Common Voice 7 test dataset for your language, we will see together how to 
+score your model - if this is the case, please don't be discouraged. We are 
+especially excited about speech recognition systems of such low-resource 
+languages and will make sure that we'll decide on a good approach to evaluating 
+your model.
+
+## Prizes
+
+TODO(Patrick, Omar, ...)
+
+## Communication and Problems
+
+If you encounter any problems or have any questions, you should use one of the following platforms
+depending on your type of problem. Hugging Face is an "open-source-first" organization meaning 
+that we'll try to solve all problems in the most public and most transparent way possible so that everybody
+in the community profits.
+
+The following table summarizes what platform to use for which problem.
+
+- Problem/question/bug with the 🤗 Datasets library that you think is a general problem that also impacts other people, please open an [Issues on Datasets](https://github.com/huggingface/datasets/issues/new?assignees=&labels=bug&template=bug-report.md&title=) and ping @anton-l and @patrickvonplaten.
+- Problem/question/bug with the 🤗 Transformers library that you think is a general problem that also impacts other people, please open an [Issues on Transformers](https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title=) and ping @anton-l and @patrickvonplaten.
+- Problem/question with a modified, customized training script that is less likely to impact other people, please post your problem/question [on the forum](https://discuss.huggingface.co/) and ping @anton-l and @patrickvonplaten.
+- Questions regarding access to the OVHcloud GPU, please ask in the Discord channel **#ovh-support**.
+- Other questions regarding the event, rules of the event, or if you are not sure where to post your question, please ask in the Discord channel **#sprint-discussions**.
+
+## Talks
+
+We are very excited to be hosting 2 days of talks from Kensho-Technologies, Mozilla's Common Voice, Kensho-Technologies, Hugging Face.
+
+### Thursday, January 20th
+
+- [Watch the talks on YouTube](TODO)
+- [Chat history](TODO)
+
+ Speaker        | Topic                           | Time                  |  Video |
+|-------------|---------------------------------|------------------------|------------------------|
+| Patrick von Platen, Hugging Face | TODO | ??? UTC      | [![Youtube](https://www.youtube.com/s/desktop/f506bd45/img/favicon_32.png)](TODO)
+| Raymond Grossman and Jeremy Lopez, Kensho-Technologies | Pyctcdecode & Speech2text decoding | 5h30pm - 6h00pm UTC      | [![Youtube](https://www.youtube.com/s/desktop/f506bd45/img/favicon_32.png)](TODO)
+
+### Friday, January 21th
+
+- [Watch the talks on YouTube](TODO)
+- [Chat history](TODO)
+
+ Speaker        | Topic                           | Time                  | Video |
+|-------------|---------------------------------|------------------------|------------------------|
+| Gabriel Habayeb, Mozilla Common Voice | TODO | 4h30pm - 5h00pm UTC      | [![Youtube](https://www.youtube.com/s/desktop/f506bd45/img/favicon_32.png)](TODO)
+| Changhan Wang, Meta AI Research | XLS-R: Large-Scale Cross-lingual Speech Representation Learning on 128 Languages | 5h30pm - 6h00pm UTC      | [![Youtube](https://www.youtube.com/s/desktop/f506bd45/img/favicon_32.png)](TODO)
+
+### Talks & Speakers
+
+#### Patrick von Platen, Research Engineer, Hugging Face
+- Talk: Introduction to Robust Speech Challenge
+- Abstract: In this talk, Patrick outlines the Robust Speech Challenge and gives tips and tricks on how to train and evaluate speech recognition systems with 🤗 Transformers and 🤗 Datasets, and PyTorch.
+- Speaker info: Patrick von Platen is a research engineer at Hugging Face and one of the core maintainers of the popular Transformers library. He specializes in speech recognition, encoder-decoder models, and long-range sequence modeling. Before joining Hugging Face, Patrick researched speech recognition at Uber AI, Cambridge University, and RWTH Aachen University.
+
+#### Raymond Grossman, Jeremy Lopez, Machine Learning Engineer, Kensho Technologies
+- Talk: PyCTCDecode & Speech2text decoding
+- Abstract: PyCTCDecode is a fast and feature-rich CTC beam search decoder for speech recognition written in Python, providing n-gram (kenlm) language model support similar to PaddlePaddle's decoder, but incorporating many new features such as byte pair encoding and real-time decoding to support models like Nvidia's Conformer-CTC or Facebook's Wav2Vec2.
+- Speaker info : 
+	- Raymond works as a machine learning engineer at Kensho Technologies, specializing in speech and natural language domains. Before coming to Kensho, he studied mathematics at Princeton and was an avid Kaggler under the moniker @ToTrainThemIsMyCause. 
+	- Jeremy is a machine learning engineer at Kensho Technologies and has worked on a variety of different topics including search and speech recognition. Before working at Kensho, he earned a PhD in experimental particle physics at MIT and continued doing physics research as a postdoc at the University of Colorado Boulder.
+
+#### Gabriel Habayeb, Data Engineer, Common Voice @ Mozilla
+- Talk: Common Voice
+- Abstract: 
+- Speaker info: 
+
+#### Changhan Wang, Main author of XLS-R and Research Engineer, Meta AI Research
+- Talk: XLS-R: Large-Scale Cross-lingual Speech Representation Learning on 128 Languages
+- Abstract: In this talk, Changhan will present XLS-R, a large-scale model for cross-lingual speech representation learning based on wav2vec 2.0. XLS-R has up to 2B parameters and was trained on nearly half a million hours of publicly available speech audio in 128 languages, an order of magnitude more public data than the largest known prior work. On the CoVoST-2 speech translation benchmark, XLS-R improves the previous state of the art by an average of 7.4 BLEU over 21 translation directions into English. For speech recognition, XLS-R improves over the best known prior work on BABEL, MLS, CommonVoice as well as VoxPopuli, lowering error rates by 14-34% relative on average. XLS-R also sets a new state of the art on VoxLingua107 language identification. The XLS-R team hopes to work together with the open-source community to improve speech processing tasks for many more languages of the world.
+
+## General Tips and Tricks
+
+- Memory efficient training:
+
+In case, you are getting out-of-memory errors on your GPU, we recommend to use 
+[bitsandbytes](https://github.com/facebookresearch/bitsandbytes) to replace the 
+native memory-intensive Adam optimizer with the one of `bitsandbytes`. You
+can simply run the script `./run_speech_recognition_ctc_bnb.py` provided in this 
+folder that makes use of `bitsandbytes` instead of the official one.
+
+- Dataset streaming
+
+TODO(Patrick)

examples/research_projects/robust-speech-event/eval.py

+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”�—’…–]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
+
+    # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # note that order is important here!
+    token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    for t in token_sequences_to_ignore:
+        text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    asr = pipeline("automatic-speech-recognition", model=args.model_id)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    args = parser.parse_args()
+
+    main(args)

examples/research_projects/robust-speech-event/run_speech_recognition_ctc_bnb.py

+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
+
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric
+
+import bitsandbytes as bnb
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_pt_utils import get_parameter_names
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.16.0.dev0")
+
+require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
+
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    activation_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
+    )
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
+            "vectors will be masked along the time axis."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    dataset_name: str = field(
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: str = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_split_name: str = field(
+        default="train+validation",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+            "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            The processor used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_length_labels (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    pad_to_multiple_of_labels: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lenghts and need
+        # different padding methods
+        input_features = [{"input_values": feature["input_values"]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.pad(
+            input_features,
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        with self.processor.as_target_processor():
+            labels_batch = self.processor.pad(
+                label_features,
+                padding=self.padding,
+                pad_to_multiple_of=self.pad_to_multiple_of_labels,
+                return_tensors="pt",
+            )
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token: Optional[str] = None,
+    unk_token: Optional[str] = None,
+    pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
+    )
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    chars_to_ignore_regex = (
+        f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
+    )
+    text_column_name = data_args.text_column_name
+
+    def remove_special_characters(batch):
+        if chars_to_ignore_regex is not None:
+            batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
+        else:
+            batch["target_text"] = batch[text_column_name].lower() + " "
+        return batch
+
+    with training_args.main_process_first(desc="dataset map special characters removal"):
+        raw_datasets = raw_datasets.map(
+            remove_special_characters,
+            remove_columns=[text_column_name],
+            desc="remove special characters from datasets",
+        )
+
+    # save special tokens for tokenizer
+    word_delimiter_token = data_args.word_delimiter_token
+    unk_token = data_args.unk_token
+    pad_token = data_args.pad_token
+
+    # 3. Next, let's load the config as we might need it to create
+    # the tokenizer
+    # load config
+    config = AutoConfig.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # 4. Next, if no tokenizer file is defined,
+    # we create the vocabulary of the model by extracting all unique characters from
+    # the training and evaluation datasets
+    # We need to make sure that only first rank saves vocabulary
+    # make sure all processes wait until vocab is created
+    tokenizer_name_or_path = model_args.tokenizer_name_or_path
+    tokenizer_kwargs = {}
+    if tokenizer_name_or_path is None:
+        # save vocab in training output dir
+        tokenizer_name_or_path = training_args.output_dir
+
+        vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+        with training_args.main_process_first():
+            if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+                os.remove(vocab_file)
+
+        with training_args.main_process_first(desc="dataset map vocabulary creation"):
+            if not os.path.isfile(vocab_file):
+                os.makedirs(tokenizer_name_or_path, exist_ok=True)
+                vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+
+                # save vocab dict to be loaded into tokenizer
+                with open(vocab_file, "w") as file:
+                    json.dump(vocab_dict, file)
+
+        # if tokenizer has just been created
+        # it is defined by `tokenizer_class` if present in config else by `model_type`
+        tokenizer_kwargs = {
+            "config": config if config.tokenizer_class is not None else None,
+            "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+            "unk_token": unk_token,
+            "pad_token": pad_token,
+            "word_delimiter_token": word_delimiter_token,
+        }
+
+    # 5. Now we can instantiate the feature extractor, tokenizer and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # derive max & min input length for sample rate & max duration
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+
+    # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+    phoneme_language = data_args.phoneme_language
+
+    # Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    def prepare_dataset(batch):
+        # load audio
+        sample = batch[audio_column_name]
+
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        batch["input_values"] = inputs.input_values[0]
+        batch["input_length"] = len(batch["input_values"])
+
+        # encode targets
+        additional_kwargs = {}
+        if phoneme_language is not None:
+            additional_kwargs["phonemizer_lang"] = phoneme_language
+
+        batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+
+    # 7. Next, we can prepare the training.
+    # Let's use word error rate (WER) as our evaluation metric,
+    # instantiate a data collator and the trainer
+
+    # Define evaluation metrics during training, *i.e.* word error rate, character error rate
+    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
+        return
+
+    def compute_metrics(pred):
+        pred_logits = pred.predictions
+        pred_ids = np.argmax(pred_logits, axis=-1)
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+        metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
+
+        return metrics
+
+    # Now save everything to be able to create a single processor later
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    try:
+        processor = AutoProcessor.from_pretrained(training_args.output_dir)
+    except (OSError, KeyError):
+        warnings.warn(
+            "Loading a processor from a feature extractor config that does not"
+            " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+            " attribute to your `preprocessor_config.json` file to suppress this warning: "
+            " `'processor_class': 'Wav2Vec2Processor'`",
+            FutureWarning,
+        )
+        processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+    decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
+    decay_parameters = [name for name in decay_parameters if "bias" not in name]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if n in decay_parameters],
+            "weight_decay": training_args.weight_decay,
+        },
+        {
+            "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
+            "weight_decay": 0.0,
+        },
+    ]
+    optimizer = bnb.optim.Adam8bit(
+        params=optimizer_grouped_parameters,
+        betas=(training_args.adam_beta1, training_args.adam_beta2),
+        eps=training_args.adam_epsilon,
+    )
+
+    optimizers = (optimizer, None)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+        optimizers=optimizers,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate()
+        max_eval_samples = (
+            data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()